fs/epoll: drop ovflist branch prediction
authorDavidlohr Bueso <dave@stgolabs.net>
Thu, 3 Jan 2019 23:27:09 +0000 (15:27 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 4 Jan 2019 21:13:46 +0000 (13:13 -0800)
commit76699a67f3041ff4c7af6d6ee9be2bfbf1ffb671
tree5d42b50be630e856b4e91c3109b56d8e4612a594
parent4e0982a00564c80cb849a892043450860ef91e14
fs/epoll: drop ovflist branch prediction

The ep->ovflist is a secondary ready-list to temporarily store events
that might occur when doing sproc without holding the ep->wq.lock.  This
accounts for every time we check for ready events and also send events
back to userspace; both callbacks, particularly the latter because of
copy_to_user, can account for a non-trivial time.

As such, the unlikely() check to see if the pointer is being used, seems
both misleading and sub-optimal.  In fact, we go to an awful lot of
trouble to sync both lists, and populating the ovflist is far from an
uncommon scenario.

For example, profiling a concurrent epoll_wait(2) benchmark, with
CONFIG_PROFILE_ANNOTATED_BRANCHES shows that for a two threads a 33%
incorrect rate was seen; and when incrementally increasing the number of
epoll instances (which is used, for example for multiple queuing load
balancing models), up to a 90% incorrect rate was seen.

Similarly, by deleting the prediction, 3% throughput boost was seen
across incremental threads.

Link: http://lkml.kernel.org/r/20181108051006.18751-4-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/eventpoll.c